HeartSaVioR commented on a change in pull request #19096: [SPARK-21869][SS] A
cached Kafka producer should not be closed if any task is using it - adds inuse
tracking.
URL: https://github.com/apache/spark/pull/19096#discussion_r272786878
##########
File path:
external/kafka-0-10-sql/src/main/scala/org/apache/spark/sql/kafka010/CachedKafkaProducer.scala
##########
@@ -40,81 +90,114 @@ private[kafka010] object CachedKafkaProducer extends
Logging {
"spark.kafka.producer.cache.timeout",
s"${defaultCacheExpireTimeout}ms")).getOrElse(defaultCacheExpireTimeout)
- private val cacheLoader = new CacheLoader[Seq[(String, Object)], Producer] {
- override def load(config: Seq[(String, Object)]): Producer = {
- val configMap = config.map(x => x._1 -> x._2).toMap.asJava
- createKafkaProducer(configMap)
+ private val cacheLoader = new CacheLoader[Seq[(String, Object)],
CachedKafkaProducer] {
+ override def load(params: Seq[(String, Object)]): CachedKafkaProducer = {
+ CachedKafkaProducer(kafkaParams = params)
}
}
- private val removalListener = new RemovalListener[Seq[(String, Object)],
Producer]() {
+ private val closeQueue = new ConcurrentLinkedQueue[CachedKafkaProducer]()
+
+ private val removalListener = new RemovalListener[Seq[(String, Object)],
CachedKafkaProducer]() {
override def onRemoval(
- notification: RemovalNotification[Seq[(String, Object)], Producer]):
Unit = {
- val paramsSeq: Seq[(String, Object)] = notification.getKey
- val producer: Producer = notification.getValue
- logDebug(
- s"Evicting kafka producer $producer params: $paramsSeq, due to
${notification.getCause}")
- close(paramsSeq, producer)
+ notification: RemovalNotification[Seq[(String, Object)],
CachedKafkaProducer]): Unit = {
+ val producer: CachedKafkaProducer = notification.getValue
+ if (producer.inUse()) {
+ logDebug(s"Evicting kafka producer $producer, due to
${notification.getCause}.")
+ // When `inuse` producer is evicted we wait for it to be released
before finally closing it.
+ closeQueue.add(producer)
+ producer.unCache()
+ } else {
+ close(producer)
+ }
}
}
- private lazy val guavaCache: LoadingCache[Seq[(String, Object)], Producer] =
+ private lazy val guavaCache: LoadingCache[Seq[(String, Object)],
CachedKafkaProducer] =
CacheBuilder.newBuilder().expireAfterAccess(cacheExpireTimeout,
TimeUnit.MILLISECONDS)
.removalListener(removalListener)
- .build[Seq[(String, Object)], Producer](cacheLoader)
-
- private def createKafkaProducer(producerConfiguration: ju.Map[String,
Object]): Producer = {
- val updatedKafkaProducerConfiguration =
- KafkaConfigUpdater("executor", producerConfiguration.asScala.toMap)
- .setAuthenticationConfigIfNeeded()
- .build()
- val kafkaProducer: Producer = new
Producer(updatedKafkaProducerConfiguration)
- logDebug(s"Created a new instance of KafkaProducer for
$updatedKafkaProducerConfiguration.")
- kafkaProducer
- }
+ .build[Seq[(String, Object)], CachedKafkaProducer](cacheLoader)
/**
* Get a cached KafkaProducer for a given configuration. If matching
KafkaProducer doesn't
* exist, a new KafkaProducer will be created. KafkaProducer is thread safe,
it is best to keep
* one instance per specified kafkaParams.
*/
- private[kafka010] def getOrCreate(kafkaParams: ju.Map[String, Object]):
Producer = {
- val paramsSeq: Seq[(String, Object)] = paramsToSeq(kafkaParams)
+ private[kafka010] def acquire(kafkaParamsMap: ju.Map[String, Object]):
CachedKafkaProducer = {
+ val paramsSeq: Seq[(String, Object)] = paramsToSeq(kafkaParamsMap)
try {
- guavaCache.get(paramsSeq)
+ val producer = this.synchronized {
+ val cachedKafkaProducer: CachedKafkaProducer =
guavaCache.get(paramsSeq)
+ cachedKafkaProducer.inUseCount.incrementAndGet()
+ logDebug(s"Granted producer $cachedKafkaProducer")
+ cachedKafkaProducer
+ }
+ producer
} catch {
- case e @ (_: ExecutionException | _: UncheckedExecutionException | _:
ExecutionError)
+ case e@(_: ExecutionException | _: UncheckedExecutionException | _:
ExecutionError)
if e.getCause != null =>
throw e.getCause
}
}
- private def paramsToSeq(kafkaParams: ju.Map[String, Object]): Seq[(String,
Object)] = {
- val paramsSeq: Seq[(String, Object)] = kafkaParams.asScala.toSeq.sortBy(x
=> x._1)
+ private def paramsToSeq(kafkaParamsMap: ju.Map[String, Object]):
Seq[(String, Object)] = {
+ val paramsSeq: Seq[(String, Object)] =
kafkaParamsMap.asScala.toSeq.sortBy(x => x._1)
paramsSeq
}
- /** For explicitly closing kafka producer */
- private[kafka010] def close(kafkaParams: ju.Map[String, Object]): Unit = {
- val paramsSeq = paramsToSeq(kafkaParams)
- guavaCache.invalidate(paramsSeq)
+ /* Release a kafka producer back to the kafka cache. We simply decrement
it's inuse count. */
+ private[kafka010] def release(producer: CachedKafkaProducer, failing:
Boolean): Unit = {
+ this.synchronized {
+ // It should be ok to call release multiple times on the same producer
object.
+ if (producer.inUse()) {
+ // So that we do not end up with -ve in-use counts.
+ producer.inUseCount.decrementAndGet()
+ logDebug(s"Released producer $producer.")
+ } else {
+ logWarning(s"Tried to release a not in use producer, $producer.")
+ }
+ if (failing) {
+ // If this producer is failing to write, we remove it from cache.
+ // So that it is re-created, eventually.
+ val cachedProducer = guavaCache.getIfPresent(producer.kafkaParams)
+ if (cachedProducer != null && cachedProducer.id == producer.id) {
+ logDebug(s"Invalidating a failing producer: $producer.")
+ guavaCache.invalidate(producer.kafkaParams)
+ }
+ }
+ }
+ if (!producer.inUse() && !producer.isCached) {
Review comment:
@ScrapCodes
> Then in such rare cases, closeQueue will hold the producer until it is
finally released by all the tasks.
If things pretty work fine, closeQueue is not needed: each task will hold
the producer when it uses the producer even closeQueue doesn't exist, and
eventually one of task will close it when decreasing inCount makes inCount
being zero.
> 1) As soon as the producer enters in the closeQueue, it cannot be acquired
by any new tasks.(Guarantee by guava)
I think it should be guaranteed by Guava when onRemoval is called with such
producer. No one can get the producer afterwards. Same without closeQueue,
right?
> 2) Once all the tasks have released it, then it is closed and removed from
the closeQueue.
closeQueue is not needed as I explained earlier.
> How closeQueue is just for mitigation of inCount goes out of sync?
As I explained race condition can happen, let's assume inCount can be out of
sync. Once producer is being added to closeQueue, even race condition I've
specified happens, the producer is added to closeQueue (though it will not be
closed at that time) and it can be closed once another producer is being
closed. So getting another chance to be closed: I meant this as "mitigation".
> How does race condition happens when isCached is true and object is not
added to closeQueue Or vice-versa?
I've already explained two times and @gaborgsomogyi can also see the case.
Please take a deep look at my comment: I understand it's not easy for author to
see defect in his/her own code.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
